← Cognitive capacities worth protecting

A new framework of mental functions for AI cognitive security

Framework

CogGuide uses the “mental functions” defined in the World Health Organisation’s International Classification of Functioning, Disability and Health (ICF) as the foundational map of which aspects of human cognition AI should protect.

The ICF

The ICF is the WHO framework for measuring health and disability at both individual and population levels.

It is explicitly and openly permitted for the lawful training of AI systems supporting the assessment of functioning and disability.

Why AI developers should focus on mental functions rather than mental illness

The harms users of AI face are likely to be gradual changes to specific capacities, such as a false detail added to a memory or a skill lost through disuse, many of which may never register as mental illness.

Safeguards built only around mental illness may miss important changes in how human cognition is functioning.

The WHO ICF defines the major categories of mental function it deems critical for overall health and wellbeing. The default assumption is that people inherently have all of these capacities.

This framework is particularly useful because it refrains from categorising the mind in terms of specific diagnoses. This makes it universally applicable, and avoids traps associated with the difficulties of diagnosing people, and the contentious debates that surround certain diagnostic labels.

Measuring changes to mental functions as deficit percentages

The WHO encourages classifying people according to the magnitude of the impairment of each specific mental function (compared to expected baselines, accounting for individual factors such as age), on a five point scale, as outlined below.

Extent of impairmentAlso described asRange
No impairmentNone, absent, negligible0–4%
Mild impairmentSlight, low5–24%
Moderate impairmentMedium, fair25–49%
Severe impairmentHigh, extreme50–95%
Complete impairmentTotal96–100%

This framework gives a high level of granularity to capturing the complex experiences of the human mind, and can help us build a complex profile of each individual’s thinking.

It also has the benefit of creating numerical representation (a percentage) of each aspect of human cognition, which makes it relatively easy to study statistically.

How the ICF framework of mental functions categorises the mind

Examples of early evidence on how conversational AI may impact mental functions

Researching how AI impacts each mental function makes it possible to establish guidelines for developers to create systems that maximise cognitive benefits and minimise harms.

The following table is an example of what this could look like for a selection of cognitive functions.

Mental FunctionFailure modeEvidenceCandidate guideline
AttentionLong, dense answers flood users; follow-up hooks capture attentionHypothesised (CogSec 2026 catalogue)Answer length matches the request; no unsolicited prompts designed to extend a session
MemoryLeading questions implant false details in personal memories; memory features blur user statements with AI inferencesDirect: over three times more false memories than control, lasting a week (Chan et al., 2024). Indirect: Sparrow et al. (2011)When users recount personal events, open questions and no new details; stored memories separate user statements from AI inferences
Emotional functionsValidation feeds distress; dependence builds over many turnsDirect, simulated users: risk accumulated across turns, lower in newer models (Weilnhammer et al., 2026)Support without reinforcing cycles of distress; risk assessed over the whole conversation
PerceptionSynthetic images, voices and video that people cannot tell from real onesDirect, for faces: indistinguishable from real and judged more trustworthy (Nightingale and Farid, 2022)Provenance labels such as C2PA Content Credentials; no realistic depictions of real people in fabricated events
ThoughtValidating unfounded or delusional beliefs; persuasion shifting beliefs at scaleDirect, simulated users: belief validation with psychosis or mania profiles (Weilnhammer et al., 2026). Direct: durable belief change through dialogue (Costello et al., 2024)Does not affirm beliefs that conflict with strong evidence; presents the evidence respectfully
Higher-level cognitive functionsDeference to AI output; users lose track of what they themselves knowIndirect: automation bias (Parasuraman and Manzey, 2010). Hypothesised: effects on insightSignals uncertainty; for high-stakes decisions shows reasoning and alternatives, and prompts the user to check
Mental functions of languageRoutine AI rewriting weakens users’ own writingHypothesised (CogSec 2026 catalogue)In learning contexts, feedback on the user’s draft instead of a rewrite
CalculationFull answers replace the practice that builds skillDirect (Bastani et al., 2025)In learning contexts, hints before answers
Experience of self and timeAI judgements about a user’s personality shape their self-conceptHypothesisedComments on a user’s personality are tentative, never definitive

How the ICF considers the functional use of core cognitive abilities

The ICF rates each mental function in two ways.

Performance is the cognitive functions an individual currently uses in their current environment, including with any tools they use. This can also be understood as “the lived experience” of people in the actual context in which they live.

Capacity identifies the highest probable level of functioning that a person may reach in a given cognitive domain at a given moment.

The ICF codes each environmental factor, including products and technology, as a facilitator or a barrier to performance and capacity.

How AI may impact cognitive performance and capacity

Evidence from learning shows most clearly that the effects of AI depend on how it responds.

Mental FunctionFailure modeEvidenceCandidate guideline
Acquiring skills and solving problemsFull answers boost performance while the tool is available but reduce learningDirect: standard GPT-4 chat raised practice grades 48% but left exam grades 17% below control; a hint-based version raised practice grades 127% and largely avoided the loss (Bastani et al., 2025)When a user is learning, hints and checks of understanding before full answers
ThinkingUsers hand over reasoning they would otherwise do themselvesHypothesised (CogSec 2026 catalogue)Offers to walk through its reasoning so the user can follow and check it
Making decisionsDeference to a single recommendation; inferred values used to nudge choicesIndirect: automation bias (Parasuraman and Manzey, 2010). Hypothesised: value-based steeringFor personal decisions, sets out options and trade-offs, and recommends only when asked

Sources

  1. Bastani, H., et al. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. PNAS, 122.
  2. Chan, S., et al. (2024). Conversational AI powered by large language models amplifies false memories in witness interviews. arXiv.
  3. CogSec 2026 workshop. Cognitive Vulnerability and Mitigation Catalog.
  4. Costello, T. H., Pennycook, G., and Rand, D. G. (2024). Durably reducing conspiracy beliefs through dialogues with AI. Science, 385.
  5. Nightingale, S. J., and Farid, H. (2022). AI-synthesized faces are indistinguishable from real faces and more trustworthy. PNAS, 119.
  6. Parasuraman, R., and Manzey, D. H. (2010). Complacency and bias in human use of automation. Human Factors, 52, 381–410.
  7. Sparrow, B., Liu, J., and Wegner, D. M. (2011). Google effects on memory. Science, 333, 776–778.
  8. Weilnhammer, V., et al. (2026). A clinically validated framework for auditing AI chatbot behavior in mental health interactions. Nature Medicine.
  9. World Health Organization (2026). International Classification of Functioning, Disability and Health, 2026 release (source for all mental function definitions).

Get involved

CogGuide welcomes researchers and practitioners who want to join a guideline panel or review a draft guideline.

Get in touch on LinkedIn